"-nan" in the fitting results of quartic IFCs

[JoeyyHuang]

Dear T. Tadano,
I am a new user of ALAMODE with latest version 1.3.0. Following the tutorial of Si example, I can get the Harmonic IFCs and phonon dispersion smoothly, then I try to fit the quartic IFCs for scph calculation. However, after set CV=19 for cross-validation automatically, I get a lot of "-nan" in the cvscore file as follow:

Algorithm : Coordinate descent
Linear model : Elastic-net
L1_RATIO = 0.8
ENET_DNORM = 1
STANDARDIZE = 1
CONV_TOL = 1e-13
95-fold cross-validation scores
L1 ALPHA, Fitting error (mean, std), Validation error (mean, std)
0.000780246 -nan -nan -nan -nan
0.000591877 -nan -nan -nan -nan
0.000448985 -nan -nan -nan -nan
0.00034059 -nan -nan -nan -nan
0.000258364 -nan -nan -nan -nan
0.000195989 -nan -nan -nan -nan
0.000148673 -nan -nan -nan -nan
0.00011278 -nan -nan -nan -nan
8.55523e-05 -nan -nan -nan -nan
6.4898e-05 -nan -nan -nan -nan
4.92302e-05 -nan -nan -nan -nan
3.73449e-05 -nan -nan -nan -nan
2.8329e-05 -nan -nan -nan -nan
2.14898e-05 -nan -nan -nan -nan
1.63016e-05 -nan -nan -nan -nan
1.23661e-05 -nan -nan -nan -nan

So I can't get the best L1 ALPHA value with the minimum value of the validation score.

I check the solution_path files, and find that for every L1 ALPHA, there is one "-nan" in the coefficients as follow:

L1 ALPHA, coefficients
0.000780246 -0 -0 -0 -0 -0.000168293 -0.000406501 0 -0 -0 -0 -0 0 -0 0 -0 0 0 0 0 -0 -0 0 -0 0 0 0 -0 0 nan 0 -0 0 0 -0 -0 -0 0 -0 0 0 0 -0 0 -0 0 0 0 0 -0 -0 0 0 0 0 0 -0 -0 0 0 -0 0 -0 0 -0 -0 0 0 -0 -0 0 0 -0 -0 -0 0 0 -0 -0 0 -0 -0 -0 -0 0 0 0
0.000591877 -0 -0 -0 -0 -0.000320923 -0.000559131 0 -0 -0 -0 -0 0 -0 0 -0 0 0 0 0 -0 -0 0 -0 0 0 0 -0 0 nan 0 -0 0 0 -0 -0 -0 0 -0 0 0 0 -0 0 -0 0 0 0 0 -0 -0 0 0 0 0 0 -0 -0 0 0 -0 0 -0 0 -0 -0 0 0 -0 -0 0 0 -0 -0 -0 0 0 -0 -0 0 -0 -0 -0 -0 0 0 0
0.000448985 -0 -0 -0 -0 -0.000436704 -0.000674913 0 -0 -0 -0 -0 0 -0 0 -0 0 0 0 0 -0 -0 0 -0 0 0 0 -0 0 nan 0 -0 0 0 -0 -0 -0 0 -0 0 0 0 -0 0 -0 0 0 0 0 -0 -0 0 0 0 0 0 -0 -0 0 0 -0 0 -0 0 -0 -0 0 0 -0 -0 0 0 -0 -0 -0 0 0 -0 -0 0 -0 -0 -0 -0 0 0 0
0.00034059 -0 -0 -0 -0 -0.000524534 -0.000762742 0 -0 -0 -0 -0 0 -0 0 -0 0 0 0 0 -0 -0 0 -0 0 0 0 -0 0 nan 0 -0 0 0 -0 -0 -0 0 -0 0 0 0 -0 0 -0 0 0 0 0 -0 -0 0 0 0 0 0 -0 -0 0 0 -0 0 -0 0 -0 -0 0 0 -0 -0 0 0 -0 -0 -0 0 0 -0 -0 0

Here is part of the input file:

&general
PREFIX = alamode
MODE = optimize
NAT = 64; NKD = 1
KD = Si
/
&interaction
NORDER = 3 # 1: harmonic, 2: cubic, ..
NBODY = 2 3 3
/
&optimize
LMODEL = enet
DFSET = DFSET_ANHARM4
NDATA = 95
FC2XML = ../alm-2/alamode.xml
CONV_TOL = 1.0e-13
MAXITER = 50000
L1_RATIO = 0.8
CV = 19
SOLUTION_PATH = 1
/
&celL
20.53404573
1.0 0.0 0.0 # a1
0.0 1.0 0.0 # a2
0.0 0.0 1.0 # a3
/
&cutoff
- None 10 6
/

The number of displacement-force training datasets(95) is larger than the number of Total Number of Free Parameters(86). I also have changed the CV value but the problem remains. Have you some ideas why this happen?

[ttadano]

Hi,

Could you try the following?

1. Use L1_RATIO = 1 (LASSO)
2. Use a smaller value for CV, say CV = 5.

[JoeyyHuang]

Dear Dr. Tadano,
Thank you for the suggestion. change the L1_RATIO and CV values do not work. But the problem is solved by enlarging the cutoff radius for quartic terms(from - None 10 6 to - None None None). Maybe no need to use so large cutoff radius, anyway, it works now.